Status CatalogManagerReplicaSet::getChunks(const BSONObj& query,
const BSONObj& sort,
boost::optional<int> limit,
vector<ChunkType>* chunks) {
chunks->clear();
auto configShard = grid.shardRegistry()->getShard("config");
auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
// Convert boost::optional<int> to boost::optional<long long>.
auto longLimit = limit ? boost::optional<long long>(*limit) : boost::none;
auto findStatus = grid.shardRegistry()->exhaustiveFind(
readHostStatus.getValue(), NamespaceString(ChunkType::ConfigNS), query, sort, longLimit);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
for (const BSONObj& obj : findStatus.getValue()) {
auto chunkRes = ChunkType::fromBSON(obj);
if (!chunkRes.isOK()) {
chunks->clear();
return {ErrorCodes::FailedToParse,
stream() << "Failed to parse chunk with id ("
<< obj[ChunkType::name()].toString()
<< "): " << chunkRes.getStatus().toString()};
}
chunks->push_back(chunkRes.getValue());
}
return Status::OK();
}
Status CatalogManagerReplicaSet::getDatabasesForShard(const string& shardName,
vector<string>* dbs) {
auto configShard = grid.shardRegistry()->getShard("config");
auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto findStatus = grid.shardRegistry()->exhaustiveFind(readHost.getValue(),
NamespaceString(DatabaseType::ConfigNS),
BSON(DatabaseType::primary(shardName)),
BSONObj(),
boost::none); // no limit
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
for (const BSONObj& obj : findStatus.getValue()) {
string dbName;
Status status = bsonExtractStringField(obj, DatabaseType::name(), &dbName);
if (!status.isOK()) {
dbs->clear();
return status;
}
dbs->push_back(dbName);
}
return Status::OK();
}
Status CatalogManagerReplicaSet::getTagsForCollection(const std::string& collectionNs,
std::vector<TagsType>* tags) {
tags->clear();
auto configShard = grid.shardRegistry()->getShard("config");
auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
auto findStatus = grid.shardRegistry()->exhaustiveFind(readHostStatus.getValue(),
NamespaceString(TagsType::ConfigNS),
BSON(TagsType::ns(collectionNs)),
BSON(TagsType::min() << 1),
boost::none); // no limit
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
for (const BSONObj& obj : findStatus.getValue()) {
auto tagRes = TagsType::fromBSON(obj);
if (!tagRes.isOK()) {
tags->clear();
return Status(ErrorCodes::FailedToParse,
str::stream()
<< "Failed to parse tag: " << tagRes.getStatus().toString());
}
tags->push_back(tagRes.getValue());
}
return Status::OK();
}
StatusWith<OpTimePair<CollectionType>> CatalogManagerReplicaSet::getCollection(
OperationContext* txn, const std::string& collNs) {
auto configShard = grid.shardRegistry()->getShard(txn, "config");
auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
auto statusFind = _exhaustiveFindOnConfig(readHostStatus.getValue(),
NamespaceString(CollectionType::ConfigNS),
BSON(CollectionType::fullNs(collNs)),
BSONObj(),
1);
if (!statusFind.isOK()) {
return statusFind.getStatus();
}
const auto& retOpTimePair = statusFind.getValue();
const auto& retVal = retOpTimePair.value;
if (retVal.empty()) {
return Status(ErrorCodes::NamespaceNotFound,
stream() << "collection " << collNs << " not found");
}
invariant(retVal.size() == 1);
auto parseStatus = CollectionType::fromBSON(retVal.front());
if (!parseStatus.isOK()) {
return parseStatus.getStatus();
}
return OpTimePair<CollectionType>(parseStatus.getValue(), retOpTimePair.opTime);
}
StatusWith<string> CatalogManagerReplicaSet::getTagForChunk(const std::string& collectionNs,
const ChunkType& chunk) {
auto configShard = grid.shardRegistry()->getShard("config");
auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
BSONObj query =
BSON(TagsType::ns(collectionNs) << TagsType::min() << BSON("$lte" << chunk.getMin())
<< TagsType::max() << BSON("$gte" << chunk.getMax()));
auto findStatus = grid.shardRegistry()->exhaustiveFind(
readHostStatus.getValue(), NamespaceString(TagsType::ConfigNS), query, BSONObj(), 1);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
const auto& docs = findStatus.getValue();
if (docs.empty()) {
return string{};
}
invariant(docs.size() == 1);
BSONObj tagsDoc = docs.front();
const auto tagsResult = TagsType::fromBSON(tagsDoc);
if (!tagsResult.isOK()) {
return {ErrorCodes::FailedToParse,
stream() << "error while parsing " << TagsType::ConfigNS << " document: " << tagsDoc
<< " : " << tagsResult.getStatus().toString()};
}
return tagsResult.getValue().getTag();
}
StatusWith<CollectionType> CatalogManagerReplicaSet::getCollection(const std::string& collNs) {
auto configShard = grid.shardRegistry()->getShard("config");
auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
auto statusFind =
grid.shardRegistry()->exhaustiveFind(readHostStatus.getValue(),
NamespaceString(CollectionType::ConfigNS),
BSON(CollectionType::fullNs(collNs)),
BSONObj(),
1);
if (!statusFind.isOK()) {
return statusFind.getStatus();
}
const auto& retVal = statusFind.getValue();
if (retVal.empty()) {
return Status(ErrorCodes::NamespaceNotFound,
stream() << "collection " << collNs << " not found");
}
invariant(retVal.size() == 1);
return CollectionType::fromBSON(retVal.front());
}
Status CatalogManagerReplicaSet::getAllShards(vector<ShardType>* shards) {
const auto configShard = grid.shardRegistry()->getShard("config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto findStatus = grid.shardRegistry()->exhaustiveFind(readHost.getValue(),
NamespaceString(ShardType::ConfigNS),
BSONObj(), // no query filter
BSONObj(), // no sort
boost::none); // no limit
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
for (const BSONObj& doc : findStatus.getValue()) {
auto shardRes = ShardType::fromBSON(doc);
if (!shardRes.isOK()) {
shards->clear();
return {ErrorCodes::FailedToParse,
stream() << "Failed to parse shard with id ("
<< doc[ShardType::name()].toString()
<< "): " << shardRes.getStatus().toString()};
}
shards->push_back(shardRes.getValue());
}
return Status::OK();
}
StatusWith<VersionType> CatalogManagerReplicaSet::_getConfigVersion(OperationContext* txn) {
const auto configShard = grid.shardRegistry()->getShard(txn, "config");
const auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
auto readHost = readHostStatus.getValue();
auto findStatus = _exhaustiveFindOnConfig(readHost,
NamespaceString(VersionType::ConfigNS),
BSONObj(),
BSONObj(),
boost::none /* no limit */);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
auto queryResults = findStatus.getValue().value;
if (queryResults.size() > 1) {
return {ErrorCodes::RemoteValidationError,
str::stream() << "should only have 1 document in " << VersionType::ConfigNS};
}
if (queryResults.empty()) {
auto countStatus =
_runCountCommandOnConfig(readHost, NamespaceString(ShardType::ConfigNS), BSONObj());
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const auto& shardCount = countStatus.getValue();
if (shardCount > 0) {
// Version document doesn't exist, but config.shards is not empty. Assuming that
// the current config metadata is pre v2.4.
VersionType versionInfo;
versionInfo.setMinCompatibleVersion(UpgradeHistory_UnreportedVersion);
versionInfo.setCurrentVersion(UpgradeHistory_UnreportedVersion);
return versionInfo;
}
VersionType versionInfo;
versionInfo.setMinCompatibleVersion(UpgradeHistory_EmptyVersion);
versionInfo.setCurrentVersion(UpgradeHistory_EmptyVersion);
return versionInfo;
}
BSONObj versionDoc = queryResults.front();
auto versionTypeResult = VersionType::fromBSON(versionDoc);
if (!versionTypeResult.isOK()) {
return Status(ErrorCodes::UnsupportedFormat,
str::stream() << "invalid config version document: " << versionDoc
<< versionTypeResult.getStatus().toString());
}
return versionTypeResult.getValue();
}
Status MigrationChunkClonerSourceLegacy::startClone(OperationContext* txn) {
invariant(!txn->lockState()->isLocked());
auto scopedGuard = MakeGuard([&] { cancelClone(txn); });
// Resolve the donor and recipient shards and their connection string
{
auto donorShardStatus = grid.shardRegistry()->getShard(txn, _args.getFromShardId());
if (!donorShardStatus.isOK()) {
return donorShardStatus.getStatus();
}
_donorCS = donorShardStatus.getValue()->getConnString();
}
{
auto recipientShardStatus = grid.shardRegistry()->getShard(txn, _args.getToShardId());
if (!recipientShardStatus.isOK()) {
return recipientShardStatus.getStatus();
}
auto recipientShard = recipientShardStatus.getValue();
auto shardHostStatus = recipientShard->getTargeter()->findHost(
ReadPreferenceSetting{ReadPreference::PrimaryOnly});
if (!shardHostStatus.isOK()) {
return shardHostStatus.getStatus();
}
_recipientHost = std::move(shardHostStatus.getValue());
}
// Prepare the currently available documents
Status status = _storeCurrentLocs(txn);
if (!status.isOK()) {
return status;
}
// Tell the recipient shard to start cloning
BSONObjBuilder cmdBuilder;
StartChunkCloneRequest::appendAsCommand(&cmdBuilder,
_args.getNss(),
_sessionId,
_args.getConfigServerCS(),
_donorCS,
_args.getFromShardId(),
_args.getToShardId(),
_args.getMinKey(),
_args.getMaxKey(),
_shardKeyPattern.toBSON(),
_args.getSecondaryThrottle());
auto responseStatus = _callRecipient(cmdBuilder.obj());
if (!responseStatus.isOK()) {
return responseStatus.getStatus();
}
scopedGuard.Dismiss();
return Status::OK();
}
void ChunkManager::calcInitSplitsAndShards(OperationContext* txn,
const ShardId& primaryShardId,
const vector<BSONObj>* initPoints,
const set<ShardId>* initShardIds,
vector<BSONObj>* splitPoints,
vector<ShardId>* shardIds) const {
verify(_chunkMap.size() == 0);
Chunk c(this,
_keyPattern.getKeyPattern().globalMin(),
_keyPattern.getKeyPattern().globalMax(),
primaryShardId);
if (!initPoints || !initPoints->size()) {
// discover split points
const auto primaryShard = grid.shardRegistry()->getShard(txn, primaryShardId);
auto targetStatus =
primaryShard->getTargeter()->findHost({ReadPreference::PrimaryPreferred, TagSet{}});
uassertStatusOK(targetStatus);
NamespaceString nss(getns());
auto result = grid.shardRegistry()->runCommand(
txn, targetStatus.getValue(), nss.db().toString(), BSON("count" << nss.coll()));
long long numObjects = 0;
uassertStatusOK(result.getStatus());
uassertStatusOK(Command::getStatusFromCommandResult(result.getValue()));
uassertStatusOK(bsonExtractIntegerField(result.getValue(), "n", &numObjects));
if (numObjects > 0)
c.pickSplitVector(txn, *splitPoints, Chunk::MaxChunkSize);
// since docs already exists, must use primary shard
shardIds->push_back(primaryShardId);
} else {
// make sure points are unique and ordered
set<BSONObj> orderedPts;
for (unsigned i = 0; i < initPoints->size(); ++i) {
BSONObj pt = (*initPoints)[i];
orderedPts.insert(pt);
}
for (set<BSONObj>::iterator it = orderedPts.begin(); it != orderedPts.end(); ++it) {
splitPoints->push_back(*it);
}
if (!initShardIds || !initShardIds->size()) {
// If not specified, only use the primary shard (note that it's not safe for mongos
// to put initial chunks on other shards without the primary mongod knowing).
shardIds->push_back(primaryShardId);
} else {
std::copy(initShardIds->begin(), initShardIds->end(), std::back_inserter(*shardIds));
}
}
}
Status AsyncRequestsSender::RemoteData::resolveShardIdToHostAndPort(
AsyncRequestsSender* ars, const ReadPreferenceSetting& readPref) {
const auto shard = getShard();
if (!shard) {
return Status(ErrorCodes::ShardNotFound,
str::stream() << "Could not find shard " << shardId);
}
auto findHostStatus = shard->getTargeter()->findHost(ars->_opCtx, readPref);
if (findHostStatus.isOK())
shardHostAndPort = std::move(findHostStatus.getValue());
return findHostStatus.getStatus();
}
Status CatalogManagerReplicaSet::getCollections(OperationContext* txn,
const std::string* dbName,
std::vector<CollectionType>* collections,
OpTime* opTime) {
BSONObjBuilder b;
if (dbName) {
invariant(!dbName->empty());
b.appendRegex(CollectionType::fullNs(),
string(str::stream() << "^" << pcrecpp::RE::QuoteMeta(*dbName) << "\\."));
}
auto configShard = grid.shardRegistry()->getShard(txn, "config");
auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto findStatus = _exhaustiveFindOnConfig(readHost.getValue(),
NamespaceString(CollectionType::ConfigNS),
b.obj(),
BSONObj(),
boost::none); // no limit
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
const auto& docsOpTimePair = findStatus.getValue();
for (const BSONObj& obj : docsOpTimePair.value) {
const auto collectionResult = CollectionType::fromBSON(obj);
if (!collectionResult.isOK()) {
collections->clear();
return {ErrorCodes::FailedToParse,
str::stream() << "error while parsing " << CollectionType::ConfigNS
<< " document: " << obj << " : "
<< collectionResult.getStatus().toString()};
}
collections->push_back(collectionResult.getValue());
}
if (opTime) {
*opTime = docsOpTimePair.opTime;
}
return Status::OK();
}
Status CatalogManagerReplicaSet::_checkDbDoesNotExist(OperationContext* txn,
const string& dbName,
DatabaseType* db) {
BSONObjBuilder queryBuilder;
queryBuilder.appendRegex(
DatabaseType::name(), (string) "^" + pcrecpp::RE::QuoteMeta(dbName) + "$", "i");
const auto configShard = grid.shardRegistry()->getShard(txn, "config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto findStatus = _exhaustiveFindOnConfig(readHost.getValue(),
NamespaceString(DatabaseType::ConfigNS),
queryBuilder.obj(),
BSONObj(),
1);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
const auto& docs = findStatus.getValue().value;
if (docs.empty()) {
return Status::OK();
}
BSONObj dbObj = docs.front();
std::string actualDbName = dbObj[DatabaseType::name()].String();
if (actualDbName == dbName) {
if (db) {
auto parseDBStatus = DatabaseType::fromBSON(dbObj);
if (!parseDBStatus.isOK()) {
return parseDBStatus.getStatus();
}
*db = parseDBStatus.getValue();
}
return Status(ErrorCodes::NamespaceExists,
str::stream() << "database " << dbName << " already exists");
}
return Status(ErrorCodes::DatabaseDifferCase,
str::stream() << "can't have 2 databases that just differ on case "
<< " have: " << actualDbName << " want to add: " << dbName);
}
StatusWith<OpTimePair<DatabaseType>> CatalogManagerReplicaSet::getDatabase(
OperationContext* txn, const std::string& dbName) {
invariant(nsIsDbOnly(dbName));
// The two databases that are hosted on the config server are config and admin
if (dbName == "config" || dbName == "admin") {
DatabaseType dbt;
dbt.setName(dbName);
dbt.setSharded(false);
dbt.setPrimary("config");
return OpTimePair<DatabaseType>(dbt);
}
const auto configShard = grid.shardRegistry()->getShard(txn, "config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto findStatus = _exhaustiveFindOnConfig(readHost.getValue(),
NamespaceString(DatabaseType::ConfigNS),
BSON(DatabaseType::name(dbName)),
BSONObj(),
1);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
const auto& docsWithOpTime = findStatus.getValue();
if (docsWithOpTime.value.empty()) {
return {ErrorCodes::DatabaseNotFound, stream() << "database " << dbName << " not found"};
}
invariant(docsWithOpTime.value.size() == 1);
auto parseStatus = DatabaseType::fromBSON(docsWithOpTime.value.front());
if (!parseStatus.isOK()) {
return parseStatus.getStatus();
}
return OpTimePair<DatabaseType>(parseStatus.getValue(), docsWithOpTime.opTime);
}
StatusWith<std::string> CatalogManagerReplicaSet::_generateNewShardName() const {
const auto configShard = grid.shardRegistry()->getShard("config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
BSONObjBuilder shardNameRegex;
shardNameRegex.appendRegex(ShardType::name(), "^shard");
auto findStatus = grid.shardRegistry()->exhaustiveFind(readHost.getValue(),
NamespaceString(ShardType::ConfigNS),
shardNameRegex.obj(),
BSON(ShardType::name() << -1),
1);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
const auto& docs = findStatus.getValue();
int count = 0;
if (!docs.empty()) {
const auto shardStatus = ShardType::fromBSON(docs.front());
if (!shardStatus.isOK()) {
return shardStatus.getStatus();
}
std::istringstream is(shardStatus.getValue().getName().substr(5));
is >> count;
count++;
}
// TODO fix so that we can have more than 10000 automatically generated shard names
if (count < 9999) {
std::stringstream ss;
ss << "shard" << std::setfill('0') << std::setw(4) << count;
return ss.str();
}
return Status(ErrorCodes::OperationFailed, "unable to generate new shard name");
}
Status AsyncResultsMerger::RemoteCursorData::resolveShardIdToHostAndPort(
const ReadPreferenceSetting& readPref) {
invariant(shardId);
invariant(!cursorId);
const auto shard = getShard();
if (!shard) {
return Status(ErrorCodes::ShardNotFound,
str::stream() << "Could not find shard " << *shardId);
}
// TODO: Pass down an OperationContext* to use here.
auto findHostStatus = shard->getTargeter()->findHostWithMaxWait(readPref, Seconds{20});
if (!findHostStatus.isOK()) {
return findHostStatus.getStatus();
}
_shardHostAndPort = std::move(findHostStatus.getValue());
return Status::OK();
}
Status AsyncResultsMerger::RemoteCursorData::resolveShardIdToHostAndPort(
OperationContext* txn, const ReadPreferenceSetting& readPref) {
invariant(shardId);
invariant(!cursorId);
const auto shard = grid.shardRegistry()->getShard(txn, *shardId);
if (!shard) {
return Status(ErrorCodes::ShardNotFound,
str::stream() << "Could not find shard " << *shardId);
}
auto findHostStatus = shard->getTargeter()->findHost(
readPref, RemoteCommandTargeter::selectFindHostMaxWaitTime(txn));
if (!findHostStatus.isOK()) {
return findHostStatus.getStatus();
}
_shardHostAndPort = std::move(findHostStatus.getValue());
return Status::OK();
}
StatusWith<SettingsType> CatalogManagerReplicaSet::getGlobalSettings(OperationContext* txn,
const string& key) {
const auto configShard = grid.shardRegistry()->getShard(txn, "config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto findStatus = _exhaustiveFindOnConfig(readHost.getValue(),
NamespaceString(SettingsType::ConfigNS),
BSON(SettingsType::key(key)),
BSONObj(),
1);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
const auto& docs = findStatus.getValue().value;
if (docs.empty()) {
return {ErrorCodes::NoMatchingDocument,
str::stream() << "can't find settings document with key: " << key};
}
BSONObj settingsDoc = docs.front();
StatusWith<SettingsType> settingsResult = SettingsType::fromBSON(settingsDoc);
if (!settingsResult.isOK()) {
return {ErrorCodes::FailedToParse,
str::stream() << "error while parsing settings document: " << settingsDoc << " : "
<< settingsResult.getStatus().toString()};
}
const SettingsType& settings = settingsResult.getValue();
Status validationStatus = settings.validate();
if (!validationStatus.isOK()) {
return validationStatus;
}
return settingsResult;
}
void MigrationManager::_executeMigrations(OperationContext* txn,
MigrationStatuses* migrationStatuses) {
for (auto& migration : _activeMigrations) {
const NamespaceString nss(migration.chunkInfo.migrateInfo.ns);
auto scopedCMStatus = ScopedChunkManager::getExisting(txn, nss);
if (!scopedCMStatus.isOK()) {
// Unable to find the ChunkManager for "nss" for whatever reason; abandon this
// migration and proceed to the next.
stdx::lock_guard<stdx::mutex> lk(_mutex);
migrationStatuses->insert(MigrationStatuses::value_type(
migration.chunkInfo.migrateInfo.getName(), std::move(scopedCMStatus.getStatus())));
continue;
}
ChunkManager* const chunkManager = scopedCMStatus.getValue().cm();
auto chunk =
chunkManager->findIntersectingChunk(txn, migration.chunkInfo.migrateInfo.minKey);
{
// No need to lock the mutex. Only this function and _takeDistLockForAMigration
// manipulate "_distributedLocks". No need to protect serial actions.
if (!_takeDistLockForAMigration(txn, migration, migrationStatuses)) {
// If there is a lock conflict between the balancer and the shard, or a shard and a
// shard, the migration has been rescheduled. Otherwise an attempt to take the lock
// failed for whatever reason and this migration is being abandoned.
continue;
}
}
const MigrationRequest& migrationRequest = migration.chunkInfo;
BSONObjBuilder builder;
MoveChunkRequest::appendAsCommand(
&builder,
nss,
chunkManager->getVersion(),
Grid::get(txn)->shardRegistry()->getConfigServerConnectionString(),
migrationRequest.migrateInfo.from,
migrationRequest.migrateInfo.to,
ChunkRange(chunk->getMin(), chunk->getMax()),
migrationRequest.maxChunkSizeBytes,
migrationRequest.secondaryThrottle,
migrationRequest.waitForDelete,
migration.oldShard ? true : false); // takeDistLock flag.
BSONObj moveChunkRequestObj = builder.obj();
const auto recipientShard =
grid.shardRegistry()->getShard(txn, migration.chunkInfo.migrateInfo.from);
const auto host = recipientShard->getTargeter()->findHost(
ReadPreferenceSetting{ReadPreference::PrimaryOnly},
RemoteCommandTargeter::selectFindHostMaxWaitTime(txn));
if (!host.isOK()) {
// Unable to find a target shard for whatever reason; abandon this migration and proceed
// to the next.
stdx::lock_guard<stdx::mutex> lk(_mutex);
migrationStatuses->insert(MigrationStatuses::value_type(
migration.chunkInfo.migrateInfo.getName(), std::move(host.getStatus())));
continue;
}
RemoteCommandRequest remoteRequest(host.getValue(), "admin", moveChunkRequestObj);
StatusWith<RemoteCommandResponse> remoteCommandResponse(
Status{ErrorCodes::InternalError, "Uninitialized value"});
executor::TaskExecutor* executor = Grid::get(txn)->getExecutorPool()->getFixedExecutor();
StatusWith<executor::TaskExecutor::CallbackHandle> callbackHandleWithStatus =
executor->scheduleRemoteCommand(remoteRequest,
stdx::bind(&MigrationManager::_checkMigrationCallback,
this,
stdx::placeholders::_1,
txn,
&migration,
migrationStatuses));
if (!callbackHandleWithStatus.isOK()) {
// Scheduling the migration moveChunk failed.
stdx::lock_guard<stdx::mutex> lk(_mutex);
migrationStatuses->insert(
MigrationStatuses::value_type(migration.chunkInfo.migrateInfo.getName(),
std::move(callbackHandleWithStatus.getStatus())));
continue;
}
// The moveChunk command was successfully scheduled. Store the callback handle so that the
// command's return can be waited for later.
stdx::lock_guard<stdx::mutex> lk(_mutex);
migration.setCallbackHandle(std::move(callbackHandleWithStatus.getValue()));
}
_waitForMigrations(txn);
// At this point, there are no parallel running threads so it is safe not to lock the mutex.
// All the migrations have returned, release all of the distributed locks that are no longer
// being used.
_distributedLocks.clear();
// If there are rescheduled migrations, move them to active and run the function again.
if (!_rescheduledMigrations.empty()) {
// Clear all the callback handles of the rescheduled migrations.
for (auto& migration : _rescheduledMigrations) {
migration.clearCallbackHandle();
}
_activeMigrations = std::move(_rescheduledMigrations);
_rescheduledMigrations.clear();
_executeMigrations(txn, migrationStatuses);
} else {
_activeMigrations.clear();
}
}
StatusWith<ShardDrainingStatus> CatalogManagerReplicaSet::removeShard(OperationContext* txn,
const std::string& name) {
const auto configShard = grid.shardRegistry()->getShard("config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
// Check preconditions for removing the shard
auto countStatus =
_runCountCommand(readHost.getValue(),
NamespaceString(ShardType::ConfigNS),
BSON(ShardType::name() << NE << name << ShardType::draining(true)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::ConflictingOperationInProgress,
"Can't have more than one draining shard at a time");
}
countStatus = _runCountCommand(readHost.getValue(),
NamespaceString(ShardType::ConfigNS),
BSON(ShardType::name() << NE << name));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() == 0) {
return Status(ErrorCodes::IllegalOperation, "Can't remove last shard");
}
// Figure out if shard is already draining
countStatus = _runCountCommand(readHost.getValue(),
NamespaceString(ShardType::ConfigNS),
BSON(ShardType::name() << name << ShardType::draining(true)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() == 0) {
log() << "going to start draining shard: " << name;
Status status = update(ShardType::ConfigNS,
BSON(ShardType::name() << name),
BSON("$set" << BSON(ShardType::draining(true))),
false, // upsert
false, // multi
NULL);
if (!status.isOK()) {
log() << "error starting removeShard: " << name << "; err: " << status.reason();
return status;
}
grid.shardRegistry()->reload();
// Record start in changelog
logChange(
txn->getClient()->clientAddress(true), "removeShard.start", "", BSON("shard" << name));
return ShardDrainingStatus::STARTED;
}
// Draining has already started, now figure out how many chunks and databases are still on the
// shard.
countStatus = _runCountCommand(
readHost.getValue(), NamespaceString(ChunkType::ConfigNS), BSON(ChunkType::shard(name)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const long long chunkCount = countStatus.getValue();
countStatus = _runCountCommand(readHost.getValue(),
NamespaceString(DatabaseType::ConfigNS),
BSON(DatabaseType::primary(name)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const long long databaseCount = countStatus.getValue();
if (chunkCount > 0 || databaseCount > 0) {
// Still more draining to do
return ShardDrainingStatus::ONGOING;
}
// Draining is done, now finish removing the shard.
log() << "going to remove shard: " << name;
audit::logRemoveShard(txn->getClient(), name);
Status status = remove(ShardType::ConfigNS, BSON(ShardType::name() << name), 0, NULL);
if (!status.isOK()) {
log() << "Error concluding removeShard operation on: " << name
<< "; err: " << status.reason();
return status;
}
grid.shardRegistry()->remove(name);
grid.shardRegistry()->reload();
// Record finish in changelog
logChange(txn->getClient()->clientAddress(true), "removeShard", "", BSON("shard" << name));
return ShardDrainingStatus::COMPLETED;
}
Status CatalogManagerReplicaSet::shardCollection(OperationContext* txn,
const string& ns,
const ShardKeyPattern& fieldsAndOrder,
bool unique,
const vector<BSONObj>& initPoints,
const set<ShardId>& initShardIds) {
// Lock the collection globally so that no other mongos can try to shard or drop the collection
// at the same time.
auto scopedDistLock = getDistLockManager()->lock(ns, "shardCollection");
if (!scopedDistLock.isOK()) {
return scopedDistLock.getStatus();
}
StatusWith<DatabaseType> status = getDatabase(nsToDatabase(ns));
if (!status.isOK()) {
return status.getStatus();
}
DatabaseType dbt = status.getValue();
ShardId dbPrimaryShardId = dbt.getPrimary();
const auto primaryShard = grid.shardRegistry()->getShard(dbPrimaryShardId);
{
// In 3.0 and prior we include this extra safety check that the collection is not getting
// sharded concurrently by two different mongos instances. It is not 100%-proof, but it
// reduces the chance that two invocations of shard collection will step on each other's
// toes. Now we take the distributed lock so going forward this check won't be necessary
// but we leave it around for compatibility with other mongoses from 3.0.
// TODO(spencer): Remove this after 3.2 ships.
const auto configShard = grid.shardRegistry()->getShard("config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto countStatus = _runCountCommand(
readHost.getValue(), NamespaceString(ChunkType::ConfigNS), BSON(ChunkType::ns(ns)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::AlreadyInitialized,
str::stream() << "collection " << ns << " already sharded with "
<< countStatus.getValue() << " chunks.");
}
}
// Record start in changelog
{
BSONObjBuilder collectionDetail;
collectionDetail.append("shardKey", fieldsAndOrder.toBSON());
collectionDetail.append("collection", ns);
collectionDetail.append("primary", primaryShard->toString());
{
BSONArrayBuilder initialShards(collectionDetail.subarrayStart("initShards"));
for (const ShardId& shardId : initShardIds) {
initialShards.append(shardId);
}
}
collectionDetail.append("numChunks", static_cast<int>(initPoints.size() + 1));
logChange(txn->getClient()->clientAddress(true),
"shardCollection.start",
ns,
collectionDetail.obj());
}
ChunkManagerPtr manager(new ChunkManager(ns, fieldsAndOrder, unique));
manager->createFirstChunks(dbPrimaryShardId, &initPoints, &initShardIds);
manager->loadExistingRanges(nullptr);
CollectionInfo collInfo;
collInfo.useChunkManager(manager);
collInfo.save(ns);
manager->reload(true);
// TODO(spencer) SERVER-19319: Send setShardVersion to primary shard so it knows to start
// rejecting unversioned writes.
BSONObj finishDetail = BSON("version"
<< ""); // TODO(spencer) SERVER-19319 Report actual version used
logChange(txn->getClient()->clientAddress(true), "shardCollection", ns, finishDetail);
return Status::OK();
}
Status CatalogManagerReplicaSet::shardCollection(OperationContext* txn,
const string& ns,
const ShardKeyPattern& fieldsAndOrder,
bool unique,
const vector<BSONObj>& initPoints,
const set<ShardId>& initShardIds) {
// Lock the collection globally so that no other mongos can try to shard or drop the collection
// at the same time.
auto scopedDistLock = getDistLockManager()->lock(ns, "shardCollection");
if (!scopedDistLock.isOK()) {
return scopedDistLock.getStatus();
}
auto status = getDatabase(txn, nsToDatabase(ns));
if (!status.isOK()) {
return status.getStatus();
}
ShardId dbPrimaryShardId = status.getValue().value.getPrimary();
const auto primaryShard = grid.shardRegistry()->getShard(txn, dbPrimaryShardId);
{
// In 3.0 and prior we include this extra safety check that the collection is not getting
// sharded concurrently by two different mongos instances. It is not 100%-proof, but it
// reduces the chance that two invocations of shard collection will step on each other's
// toes. Now we take the distributed lock so going forward this check won't be necessary
// but we leave it around for compatibility with other mongoses from 3.0.
// TODO(spencer): Remove this after 3.2 ships.
const auto configShard = grid.shardRegistry()->getShard(txn, "config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto countStatus = _runCountCommandOnConfig(
readHost.getValue(), NamespaceString(ChunkType::ConfigNS), BSON(ChunkType::ns(ns)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::AlreadyInitialized,
str::stream() << "collection " << ns << " already sharded with "
<< countStatus.getValue() << " chunks.");
}
}
// Record start in changelog
{
BSONObjBuilder collectionDetail;
collectionDetail.append("shardKey", fieldsAndOrder.toBSON());
collectionDetail.append("collection", ns);
collectionDetail.append("primary", primaryShard->toString());
{
BSONArrayBuilder initialShards(collectionDetail.subarrayStart("initShards"));
for (const ShardId& shardId : initShardIds) {
initialShards.append(shardId);
}
}
collectionDetail.append("numChunks", static_cast<int>(initPoints.size() + 1));
logChange(txn,
txn->getClient()->clientAddress(true),
"shardCollection.start",
ns,
collectionDetail.obj());
}
shared_ptr<ChunkManager> manager(new ChunkManager(ns, fieldsAndOrder, unique));
manager->createFirstChunks(txn, dbPrimaryShardId, &initPoints, &initShardIds);
manager->loadExistingRanges(txn, nullptr);
CollectionInfo collInfo;
collInfo.useChunkManager(manager);
collInfo.save(txn, ns);
manager->reload(txn, true);
// Tell the primary mongod to refresh its data
// TODO: Think the real fix here is for mongos to just
// assume that all collections are sharded, when we get there
SetShardVersionRequest ssv = SetShardVersionRequest::makeForVersioningNoPersist(
grid.shardRegistry()->getConfigServerConnectionString(),
dbPrimaryShardId,
primaryShard->getConnString(),
NamespaceString(ns),
ChunkVersionAndOpTime(manager->getVersion(), manager->getConfigOpTime()),
true);
auto ssvStatus = grid.shardRegistry()->runCommandWithNotMasterRetries(
txn, dbPrimaryShardId, "admin", ssv.toBSON());
if (!ssvStatus.isOK()) {
warning() << "could not update initial version of " << ns << " on shard primary "
<< dbPrimaryShardId << ssvStatus.getStatus();
}
logChange(txn,
txn->getClient()->clientAddress(true),
"shardCollection",
ns,
BSON("version" << manager->getVersion().toString()));
return Status::OK();
}
void AsyncResultsMerger::handleBatchResponse(
const executor::TaskExecutor::RemoteCommandCallbackArgs& cbData, size_t remoteIndex) {
stdx::lock_guard<stdx::mutex> lk(_mutex);
auto& remote = _remotes[remoteIndex];
// Clear the callback handle. This indicates that we are no longer waiting on a response from
// 'remote'.
remote.cbHandle = executor::TaskExecutor::CallbackHandle();
// If we're in the process of shutting down then there's no need to process the batch.
if (_lifecycleState != kAlive) {
invariant(_lifecycleState == kKillStarted);
// Make sure to wake up anyone waiting on '_currentEvent' if we're shutting down.
signalCurrentEventIfReady_inlock();
// If we're killed and we're not waiting on any more batches to come back, then we are ready
// to kill the cursors on the remote hosts and clean up this cursor. Schedule the
// killCursors command and signal that this cursor is safe now safe to destroy. We have to
// promise not to touch any members of this class because 'this' could become invalid as
// soon as we signal the event.
if (!haveOutstandingBatchRequests_inlock()) {
// If the event handle is invalid, then the executor is in the middle of shutting down,
// and we can't schedule any more work for it to complete.
if (_killCursorsScheduledEvent.isValid()) {
scheduleKillCursors_inlock();
_executor->signalEvent(_killCursorsScheduledEvent);
}
_lifecycleState = kKillComplete;
}
return;
}
// Early return from this point on signal anyone waiting on an event, if ready() is true.
ScopeGuard signaller = MakeGuard(&AsyncResultsMerger::signalCurrentEventIfReady_inlock, this);
if (!cbData.response.isOK()) {
remote.status = cbData.response.getStatus();
// If we failed to retrieve the batch because we couldn't contact the remote, we notify that
// targeter that the host is unreachable. The caller can then retry on a new host.
if (remote.status == ErrorCodes::HostUnreachable && remote.shardId) {
auto shard = _params.shardRegistry->getShard(_params.txn, *remote.shardId);
if (!shard) {
remote.status =
Status(ErrorCodes::HostUnreachable,
str::stream() << "Could not find shard " << *remote.shardId
<< " containing host " << remote.hostAndPort.toString());
} else {
shard->getTargeter()->markHostUnreachable(remote.hostAndPort);
}
}
return;
}
auto getMoreParseStatus = CursorResponse::parseFromBSON(cbData.response.getValue().data);
if (!getMoreParseStatus.isOK()) {
remote.status = getMoreParseStatus.getStatus();
return;
}
auto cursorResponse = getMoreParseStatus.getValue();
// If we have a cursor established, and we get a non-zero cursorid that is not equal to the
// established cursorid, we will fail the operation.
if (remote.cursorId && cursorResponse.cursorId != 0 &&
*remote.cursorId != cursorResponse.cursorId) {
remote.status = Status(ErrorCodes::BadValue,
str::stream() << "Expected cursorid " << *remote.cursorId
<< " but received " << cursorResponse.cursorId);
return;
}
remote.cursorId = cursorResponse.cursorId;
remote.cmdObj = boost::none;
for (const auto& obj : cursorResponse.batch) {
// If there's a sort, we're expecting the remote node to give us back a sort key.
if (!_params.sort.isEmpty() &&
obj[ClusterClientCursorParams::kSortKeyField].type() != BSONType::Object) {
remote.status = Status(ErrorCodes::InternalError,
str::stream() << "Missing field '"
<< ClusterClientCursorParams::kSortKeyField
<< "' in document: " << obj);
return;
}
remote.docBuffer.push(obj);
++remote.fetchedCount;
}
// If we're doing a sorted merge, then we have to make sure to put this remote onto the
// merge queue.
if (!_params.sort.isEmpty() && !cursorResponse.batch.empty()) {
_mergeQueue.push(remoteIndex);
}
// If the cursor is tailable and we just received an empty batch, the next return value should
// be boost::none in order to indicate the end of the batch.
if (_params.isTailable && !remote.hasNext()) {
_eofNext = true;
}
// If even after receiving this batch we still don't have anything buffered (i.e. the batchSize
// was zero), then can schedule work to retrieve the next batch right away.
//
// We do not ask for the next batch if the cursor is tailable, as batches received from remote
// tailable cursors should be passed through to the client without asking for more batches.
if (!_params.isTailable && !remote.hasNext() && !remote.exhausted()) {
auto nextBatchStatus = askForNextBatch_inlock(remoteIndex);
if (!nextBatchStatus.isOK()) {
remote.status = nextBatchStatus;
return;
}
}
// ScopeGuard requires dismiss on success, but we want waiter to be signalled on success as
// well as failure.
signaller.Dismiss();
signalCurrentEventIfReady_inlock();
}
StatusWith<VersionType> CatalogManagerReplicaSet::_getConfigVersion() {
const auto configShard = grid.shardRegistry()->getShard("config");
const auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
auto readHost = readHostStatus.getValue();
auto findStatus = grid.shardRegistry()->exhaustiveFind(readHost,
NamespaceString(VersionType::ConfigNS),
BSONObj(),
BSONObj(),
boost::none /* no limit */);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
auto queryResults = findStatus.getValue();
if (queryResults.size() > 1) {
return {ErrorCodes::RemoteValidationError,
str::stream() << "should only have 1 document in " << VersionType::ConfigNS};
}
if (queryResults.empty()) {
auto cmdStatus =
grid.shardRegistry()->runCommand(readHost, "admin", BSON("listDatabases" << 1));
if (!cmdStatus.isOK()) {
return cmdStatus.getStatus();
}
const BSONObj& cmdResult = cmdStatus.getValue();
Status cmdResultStatus = getStatusFromCommandResult(cmdResult);
if (!cmdResultStatus.isOK()) {
return cmdResultStatus;
}
for (const auto& dbEntry : cmdResult["databases"].Obj()) {
const string& dbName = dbEntry["name"].String();
if (dbName != "local" && dbName != "admin") {
VersionType versionInfo;
versionInfo.setMinCompatibleVersion(UpgradeHistory_UnreportedVersion);
versionInfo.setCurrentVersion(UpgradeHistory_UnreportedVersion);
return versionInfo;
}
}
VersionType versionInfo;
versionInfo.setMinCompatibleVersion(UpgradeHistory_EmptyVersion);
versionInfo.setCurrentVersion(UpgradeHistory_EmptyVersion);
return versionInfo;
}
BSONObj versionDoc = queryResults.front();
auto versionTypeResult = VersionType::fromBSON(versionDoc);
if (!versionTypeResult.isOK()) {
return Status(ErrorCodes::UnsupportedFormat,
str::stream() << "invalid config version document: " << versionDoc
<< versionTypeResult.getStatus().toString());
}
return versionTypeResult.getValue();
}
void AsyncResultsMerger::handleBatchResponse(
const executor::TaskExecutor::RemoteCommandCallbackArgs& cbData, size_t remoteIndex) {
stdx::lock_guard<stdx::mutex> lk(_mutex);
auto& remote = _remotes[remoteIndex];
// Clear the callback handle. This indicates that we are no longer waiting on a response from
// 'remote'.
remote.cbHandle = executor::TaskExecutor::CallbackHandle();
// If we're in the process of shutting down then there's no need to process the batch.
if (_lifecycleState != kAlive) {
invariant(_lifecycleState == kKillStarted);
// Make sure to wake up anyone waiting on '_currentEvent' if we're shutting down.
signalCurrentEventIfReady_inlock();
// Make a best effort to parse the response and retrieve the cursor id. We need the cursor
// id in order to issue a killCursors command against it.
if (cbData.response.isOK()) {
auto cursorResponse = parseCursorResponse(cbData.response.getValue().data, remote);
if (cursorResponse.isOK()) {
remote.cursorId = cursorResponse.getValue().getCursorId();
}
}
// If we're killed and we're not waiting on any more batches to come back, then we are ready
// to kill the cursors on the remote hosts and clean up this cursor. Schedule the
// killCursors command and signal that this cursor is safe now safe to destroy. We have to
// promise not to touch any members of this class because 'this' could become invalid as
// soon as we signal the event.
if (!haveOutstandingBatchRequests_inlock()) {
// If the event handle is invalid, then the executor is in the middle of shutting down,
// and we can't schedule any more work for it to complete.
if (_killCursorsScheduledEvent.isValid()) {
scheduleKillCursors_inlock();
_executor->signalEvent(_killCursorsScheduledEvent);
}
_lifecycleState = kKillComplete;
}
return;
}
// Early return from this point on signal anyone waiting on an event, if ready() is true.
ScopeGuard signaller = MakeGuard(&AsyncResultsMerger::signalCurrentEventIfReady_inlock, this);
StatusWith<CursorResponse> cursorResponseStatus(
cbData.response.isOK() ? parseCursorResponse(cbData.response.getValue().data, remote)
: cbData.response.getStatus());
if (!cursorResponseStatus.isOK()) {
// Notify the shard registry of the failure.
if (remote.shardId) {
auto shard = grid.shardRegistry()->getShard(_params.txn, *remote.shardId);
if (!shard) {
remote.status = Status(cursorResponseStatus.getStatus().code(),
str::stream() << "Could not find shard " << *remote.shardId
<< " containing host "
<< remote.getTargetHost().toString());
} else {
ShardRegistry::updateReplSetMonitor(
shard->getTargeter(), remote.getTargetHost(), cursorResponseStatus.getStatus());
}
}
// If the error is retriable, schedule another request.
if (!remote.cursorId && remote.retryCount < kMaxNumFailedHostRetryAttempts &&
isPerShardRetriableError(cursorResponseStatus.getStatus().code())) {
++remote.retryCount;
// Since we potentially updated the targeter that the last host it chose might be
// faulty, the call below may end up getting a different host.
remote.status = askForNextBatch_inlock(remoteIndex);
if (remote.status.isOK()) {
return;
}
// If we end up here, it means we failed to schedule the retry request, which is a more
// severe error that should not be retried. Just pass through to the error handling
// logic below.
} else {
remote.status = cursorResponseStatus.getStatus();
}
// Unreachable host errors are swallowed if the 'allowPartialResults' option is set. We
// remove the unreachable host entirely from consideration by marking it as exhausted.
if (_params.isAllowPartialResults) {
remote.status = Status::OK();
// Clear the results buffer and cursor id.
std::queue<BSONObj> emptyBuffer;
std::swap(remote.docBuffer, emptyBuffer);
remote.cursorId = 0;
}
return;
}
// Cursor id successfully established.
auto cursorResponse = std::move(cursorResponseStatus.getValue());
remote.cursorId = cursorResponse.getCursorId();
remote.initialCmdObj = boost::none;
for (const auto& obj : cursorResponse.getBatch()) {
// If there's a sort, we're expecting the remote node to give us back a sort key.
if (!_params.sort.isEmpty() &&
obj[ClusterClientCursorParams::kSortKeyField].type() != BSONType::Object) {
remote.status = Status(ErrorCodes::InternalError,
str::stream() << "Missing field '"
<< ClusterClientCursorParams::kSortKeyField
<< "' in document: " << obj);
return;
}
remote.docBuffer.push(obj);
++remote.fetchedCount;
}
// If we're doing a sorted merge, then we have to make sure to put this remote onto the
// merge queue.
if (!_params.sort.isEmpty() && !cursorResponse.getBatch().empty()) {
_mergeQueue.push(remoteIndex);
}
// If the cursor is tailable and we just received an empty batch, the next return value should
// be boost::none in order to indicate the end of the batch.
if (_params.isTailable && !remote.hasNext()) {
_eofNext = true;
}
// If even after receiving this batch we still don't have anything buffered (i.e. the batchSize
// was zero), then can schedule work to retrieve the next batch right away.
//
// We do not ask for the next batch if the cursor is tailable, as batches received from remote
// tailable cursors should be passed through to the client without asking for more batches.
if (!_params.isTailable && !remote.hasNext() && !remote.exhausted()) {
remote.status = askForNextBatch_inlock(remoteIndex);
if (!remote.status.isOK()) {
return;
}
}
// ScopeGuard requires dismiss on success, but we want waiter to be signalled on success as
// well as failure.
signaller.Dismiss();
signalCurrentEventIfReady_inlock();
}